Brandy,
DuGoff et al (2014) have a great section in their paper where they describe 28 studies with some kind of survey weights, that use propensity score methods, and yet approach the problem in a variety of ways.
- 16 studies ignored the weights completely (while claiming representativeness)
- 7 used the weights only in the outcome model
- 5 used the weights in both the propensity score and outcome model
Clearly there's confusion on how to handle survey weights with propensity score analysis. While it doesn't address complex survey design generally (e.g. cluster sampling), some colleagues and I showed mathematically how you should incorporate sampling weights into a propensity score analysis.
G. Ridgeway, S. Kovalchik, B.A. Griffin, and M.U. Kabeto (2015). "Propensity score analysis with survey weighted data," Journal of Causal Inference 3(2):237-249.
The primary conclusion is that you should use sampling weights in the propensity score estimation stage (as weights, not as a covariate), compute final weights as the product of the sampling weight and the propensity score weight, and use those final weights in an outcome model.
You will find papers that claim that you do not need to use the sampling weights in the propensity score estimation stage. You will also note that those same papers have no mathematics supporting their claims... just simulation or a particular applied example in which it turns out not to matter. We point out three specific cases that will cause problems if you do not use the sampling weights in the propensity score estimation stage:
- If there is a covariate <inline-formula><alternatives>z</alternatives></inline-formula> used in the sampling weights that is not used or available for the propensity score model even if <inline-formula><alternatives>z</alternatives></inline-formula> is independent of the potential outcomes
- If the propensity score model has limited degrees of freedom and spends those degrees of freedom on the domain of pretreatment covariates x with small sampling weights
- If the sampling probability depends on treatment assignment, particularly for the case when treatment and control cases are drawn from different survey efforts or different survey waves
David Lenis has some recent papers too that address other specific issues
https://doi.org/10.1016/j.csda.2018.05.003https://doi.org/10.1093/biostatistics/kxx063Greg
------------------------------
Greg Ridgeway
Professor and Chair, Department of Criminology
Professor, Department of Statistics
University of Pennsylvania
------------------------------
Original Message:
Sent: 09-16-2020 09:53
From: Brandy Sinco
Subject: Propensity score matching on Complex Survey Data
Hi ASA Colleagues:
Does anyone have recommendations for articles on propensity score matching with complex survey data?
These are on my current reading list and I'd appreciate suggestions for other articles.
Austin, Jembere, and Chiu (2018). Propensity score matching and complex surveys.
DuGoff, Schuler, Stuart (2014). Generalizing Observational Study Results: Applying Propensity Score Methods to Complex Surveys.
Karabon (2019). Applying Propensity Score Methods to Complex Survey Data Using SAS PROC PSMATCH.
------------------------------
Brandy Sinco, BS, MA, MS
Statistician Senior
Michigan Medicine
------------------------------